Cohue

Inspiration

Most good ideas die between "someone should build this" and an actual team. Any crowd (a hackathon hall, a campus, a neighborhood) has people whose skills fit together, but nobody can see them. Availability is a guessing game, language shuts people out, and "maybe" isn't a commitment. AI assistants give advice, directories list profiles and project tools organize teams that already exist. Nothing handles the step in between: turning strangers into a team that has actually agreed to work together.

Cohue is French for a bustling crowd. We built COHUE to turn that crowd into purposeful, agreed action, without ever deciding for anyone.

What it does

COHUE turns a goal into a real, agreed team, across languages. Try it at https://cohue.tech.

Describe a mission, typed or spoken. The Coordinator chats with you in your language (and can answer out loud), then drafts the skills needed, group size and time window. A second AI, the Strategist, adds a reality check such as "15 minutes is unrealistic for a 10-person repair café". You review and confirm every field.
Deterministic matching. COHUE searches people who opted in and are discoverable, and proposes groups that cover every required skill and share enough time. It explains coverage, overlap and ranking for each group. If no team is possible, it says so instead of faking one.
Consent on each person's own phone. The organizer shares a personal link or QR code. Each invitee reads the full terms (role, responsibilities, schedule, teammates) and accepts, declines or suggests another time. A suggestion is never a yes, and no AI can accept for anyone.
Repair and reconsent. A decline triggers a re-plan. If the team changes, everyone approves the new terms again.
Activation and Mission Room. Only when every member has approved does the mission activate, creating memberships, first tasks and a shared brief in one step. The Room has a task board, the team with usernames and avatars, the brief in each member's language, an AI first-hour kickoff plan for each role, and public resources marked unverified.
Multilingual. Invitations and briefs are translated into English, Spanish or French, marked "review required" until the organizer approves them, and can be played as speech. The binding original is always shown.
Live mission theater. A network map, an agent and tool activity board and an activity feed update only from saved events. You can replay any past moment, and every view has reduced-motion and list alternatives.

How we built it

App: Next.js 16, React 19 and TypeScript as one modular app, with shared Zod contracts for every API call and event. Tailwind 4, Radix UI, React Flow + ELK, and Motion/GSAP animations triggered only by committed events.
Data: PostgreSQL on Tiger Data with Drizzle. Every command is idempotent, versioned and scoped to live or synthetic data.
Durable worker: a leased, fenced job runner on a Vultr VM (systemd), so missions keep progressing with every browser closed.
Matching: a custom deterministic constraint search over skills, schedule overlap and capacity.
Google Gemini: drafts structured requirements, profiles and translations. Its output is always validated and reviewed, and never trusted for consent or feasibility.
Meta Model API (Muse Spark): the Strategist's second opinion and a hedged fallback. If Gemini hasn't answered in 6 seconds, Meta runs in parallel and the first usable answer wins.
ElevenLabs: speech-to-text (Scribe) and the Coordinator's voice and playback (Multilingual v2).
SerpApi: public resource discovery, shown only as unverified evidence and never as a team member.
Vercel hosting at cohue.tech, plus GitHub Actions CI running Vitest and Playwright against a real PostgreSQL database.

Challenges we ran into

Keeping consent truly human. A counterproposal must trigger repair without counting as acceptance, a changed team must invalidate earlier approvals, and two people accepting overlapping missions at the same instant must never be double-booked.
An ordering bug. Integration testing caught revocations for old invitations arriving after their replacements, which made new invitations look revoked. Committing revocations before issuing replacements fixed it.
Proving the worker survives crashes. We added leases and fencing so a crashed worker can never overwrite its replacement's work.
Translating without changing the deal. A model must never quietly change a time or a responsibility.
Live-demo latency. One slow model could stall the composer, so we hedged across two providers.

Accomplishments that we're proud of

A real two-phone run on the live site. An English organizer, a Spanish-speaking designer and a French-speaking hardware builder formed a team, accepted on their own phones and activated a mission with briefs in all three languages.
The crash test on production. With every browser closed, we killed the worker mid-job on the Vultr host. It restarted, recovered the job after the lease expired, the dead worker's late write was rejected, and exactly one activation happened.
Matching at scale. On a 200-person synthetic crowd, a 3-person mission checked 4,372 team combinations in 14 ms and returned 5 valid proposals.
Automated end-to-end tests cover decline, repair, reconsent, concurrent acceptance and activation against a real database, run in CI on every push.

What we learned

Coordination needs a precise record of who agreed to which version of a plan. Making the "boring" guarantees explicit early (idempotency, versions, scope, consent hashes) is what made the AI parts safe to add. A convincing animation should follow a real, committed change, so people can see what happened and what still needs their decision.

What's next for COHUE

More languages after validation, importing availability from calendars, richer Mission Rooms, and wider real-device accessibility testing. After that, we want to pilot COHUE with student clubs and community projects, where complementary skills and a shared goal can become a team.

Notes for judges

At https://cohue.tech, choose Explore the synthetic showcase. The 200 showcase people are clearly labeled synthetic, and their responses are labeled operator decisions. Matching, consent, repair and activation run on the same real engine as live missions. Source: https://github.com/richmondazadze/cohue

AI tools disclosure

Claude Code and Codex assisted with implementation, debugging, documentation and testing. Real people made the mission and consent decisions in the recorded live run.